Classifying Upper Limb Activities Using Deep Neural Networks

This paper presents a classification method using Inertial Measurement Unit (IMU) in order to classify six human upper limb activities. The study was also carried out to investigate whether theses activities are being performed normally or abnormally using two different neural networks: Artificial neural network (ANN) and convolutional neural network (CNN). Human activities that were included in the study: arm flexion and extension, arm pronation and supination, shoulder internal and external rotations. Before activities categorization, training data was obtained by the means of an IMU sensor fixed on an armband worn around the forearm. The training data obtained were positions, velocities, accelerations and jerks around x, y and z axes. Training samples of 264 have been collected from 10 participants, 2 women and 8 men from ages 19 to 23. Then, 204 features were extracted from IMU data, nonetheless, 15 features only have been used as inputs to the proposed neural networks because they were the most distinguished ones. After all, the networks classify the data into one of 6 classes and their results were compared. Furthermore, these proposed methods of classification have been validated by real experiments showing that ANN network gives the best performance. © 2020, Springer Nature Switzerland AG.